drm/amdgpu: optimize scheduler fence handling

We only need to wait for jobs to be scheduled when
the dependency is from the same scheduler.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>

diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
index ea30d6a..b7cd108 100644 (file)
--- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
+++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.c
@@ -211,6 +211,41 @@ static void amd_sched_entity_wakeup(struct fence *f, struct fence_cb *cb)
        amd_sched_wakeup(entity->sched);
 }
 
+static bool amd_sched_entity_add_dependency_cb(struct amd_sched_entity *entity)
+{
+       struct amd_gpu_scheduler *sched = entity->sched;
+       struct fence * fence = entity->dependency;
+       struct amd_sched_fence *s_fence;
+
+       if (fence->context == entity->fence_context) {
+               /* We can ignore fences from ourself */
+               fence_put(entity->dependency);
+               return false;
+       }
+
+       s_fence = to_amd_sched_fence(fence);
+       if (s_fence && s_fence->sched == sched) {
+               /* Fence is from the same scheduler */
+               if (test_bit(AMD_SCHED_FENCE_SCHEDULED_BIT, &fence->flags)) {
+                       /* Ignore it when it is already scheduled */
+                       fence_put(entity->dependency);
+                       return false;
+               }
+
+               /* Wait for fence to be scheduled */
+               entity->cb.func = amd_sched_entity_wakeup;
+               list_add_tail(&entity->cb.node, &s_fence->scheduled_cb);
+               return true;
+       }
+
+       if (!fence_add_callback(entity->dependency, &entity->cb,
+                               amd_sched_entity_wakeup))
+               return true;
+
+       fence_put(entity->dependency);
+       return false;
+}
+
 static struct amd_sched_job *
 amd_sched_entity_pop_job(struct amd_sched_entity *entity)
 {
@@ -223,20 +258,9 @@ amd_sched_entity_pop_job(struct amd_sched_entity *entity)
        if (!kfifo_out_peek(&entity->job_queue, &sched_job, sizeof(sched_job)))
                return NULL;
 
-       while ((entity->dependency = sched->ops->dependency(sched_job))) {
-
-               if (entity->dependency->context == entity->fence_context) {
-                       /* We can ignore fences from ourself */
-                       fence_put(entity->dependency);
-                       continue;
-               }
-
-               if (fence_add_callback(entity->dependency, &entity->cb,
-                                      amd_sched_entity_wakeup))
-                       fence_put(entity->dependency);
-               else
+       while ((entity->dependency = sched->ops->dependency(sched_job)))
+               if (amd_sched_entity_add_dependency_cb(entity))
                        return NULL;
-       }
 
        return sched_job;
 }
@@ -400,6 +424,7 @@ static int amd_sched_main(void *param)
 
                atomic_inc(&sched->hw_rq_count);
                fence = sched->ops->run_job(sched_job);
+               amd_sched_fence_scheduled(s_fence);
                if (fence) {
                        r = fence_add_callback(fence, &s_fence->cb,
                                               amd_sched_process_job);

diff --git a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
index 939692b..a0f0ae5 100644 (file)
--- a/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
+++ b/drivers/gpu/drm/amd/scheduler/gpu_scheduler.h
@@ -27,6 +27,8 @@
 #include <linux/kfifo.h>
 #include <linux/fence.h>
 
+#define AMD_SCHED_FENCE_SCHEDULED_BIT  FENCE_FLAG_USER_BITS
+
 struct amd_gpu_scheduler;
 struct amd_sched_rq;
 
@@ -68,6 +70,7 @@ struct amd_sched_rq {
 struct amd_sched_fence {
        struct fence                    base;
        struct fence_cb                 cb;
+       struct list_head                scheduled_cb;
        struct amd_gpu_scheduler        *sched;
        spinlock_t                      lock;
        void                            *owner;
@@ -134,7 +137,7 @@ void amd_sched_entity_push_job(struct amd_sched_job *sched_job);
 
 struct amd_sched_fence *amd_sched_fence_create(
        struct amd_sched_entity *s_entity, void *owner);
+void amd_sched_fence_scheduled(struct amd_sched_fence *fence);
 void amd_sched_fence_signal(struct amd_sched_fence *fence);
 
-
 #endif

diff --git a/drivers/gpu/drm/amd/scheduler/sched_fence.c b/drivers/gpu/drm/amd/scheduler/sched_fence.c
index 8d2130b..87c78ee 100644 (file)
--- a/drivers/gpu/drm/amd/scheduler/sched_fence.c
+++ b/drivers/gpu/drm/amd/scheduler/sched_fence.c
@@ -35,6 +35,8 @@ struct amd_sched_fence *amd_sched_fence_create(struct amd_sched_entity *s_entity
        fence = kmem_cache_zalloc(sched_fence_slab, GFP_KERNEL);
        if (fence == NULL)
                return NULL;
+
+       INIT_LIST_HEAD(&fence->scheduled_cb);
        fence->owner = owner;
        fence->sched = s_entity->sched;
        spin_lock_init(&fence->lock);
@@ -55,6 +57,17 @@ void amd_sched_fence_signal(struct amd_sched_fence *fence)
                FENCE_TRACE(&fence->base, "was already signaled\n");
 }
 
+void amd_sched_fence_scheduled(struct amd_sched_fence *s_fence)
+{
+       struct fence_cb *cur, *tmp;
+
+       set_bit(AMD_SCHED_FENCE_SCHEDULED_BIT, &s_fence->base.flags);
+       list_for_each_entry_safe(cur, tmp, &s_fence->scheduled_cb, node) {
+               list_del_init(&cur->node);
+               cur->func(&s_fence->base, cur);
+       }
+}
+
 static const char *amd_sched_fence_get_driver_name(struct fence *fence)
 {
        return "amd_sched";